When will we start seeing updates to the software that will tighten the detection method? Are there being steps taken to help reduce the false positives that customer regularly get? Are we going to see any further updates to the mobile application? Will there be any steps towards facial recognition?
I like your product a lot, but a quick review of posts shows that there are a lot of issues around Bug Detection. Other companies are making advancements to determine the difference between a human and a bug on the screen.
Best way to reduce false from bugs is still to have a separate IR source. Mobile App was updated in April and is working fine. Facial recognition is there, but I have never used it. What other companies?
-Henrik
I have problems with false detections as I guess everybody has. While more advanced solutions like suggested in this thread might solve the issue I have tried to tweak setting for months but still get false positives for (fast) flying bugs, moving tree branch shadows and even clouds changing the light conditions of the entire scene.
I find it very hard to tweak the settings to be better as there is no way (that I know of) to take a recording that produced a false detection and run it again with new tweaked settings. Only way is to tweak settings and then wait for days/weeks trying to guess if the tweak was good or not. And the possibility to filter out real events that I want to record is also possible with aggressive tweaks. Not good.
If it was possible to have some sort of self-recorded library with video streams and then have the possibility to run them through the motion detection algoritm it would be quite easy to tweak settings for my camera and environment. The entire library could be automatically tested for true/false detections as well!
Does this sound like a good idea for a future version or is it already possible to achieve or similar functionality in some way in order to tweak setting in a more controlled way?
Hi and thanks for your comments.
I certainly agree that this is a very interesting area to look into an improve. With high resolution camera and the new i9 generation processors one can certainly do a lot of interesting calculations. Learning the system whatās false, correlation techniques etc.
I agree it is good ideas for a future version. If that will be in a free version can be discussed since it will need āsomeā development also for the exiting price. Well, itĀ“s on the table now ā¦
i know this thread is 4yrs old, though seems most relevantā¦ for those interested in motion detection v. what otherwise is essentially semi-continuous recording:
While Features.MotionDetectin.DetectionZones can be useful (aside from periodic camera resets where it fails to recall prior position, resulting in need for manual repositioning and then potentially tweaking if exact prior position isnāt obtained)ā¦ itās often still problematic throughout the day as sun shifts and shaddows change, let alone wind causing moving tree shadows, and at dusk/dawn changes, and at night with IR and movement noise on background areas, etc and thus amounts to semi-continuous recording.
With all thatās gone on of late with Machine Learning, seems like in conjunction w/Zones or not, it would be great to add trained object detection & user trainingā¦ so that waving branches, and shadows, and grass ānoiseā at night, and moths in IR are ignored, and only people, and/or autos, and perhaps cats or dogs are detected and nothing else.
And maybe (per the initial inquiry) you allow folks to train their own valid variations-- eg. just Fedex/UPS/Amazon/USPS delivery trucks v. other vehicles, for delivery notifications.
Thereās great reusable stuff on this front ā eg. [Train Object Detection AI with 6 lines of code | by Moses Olafenwa | DeepQuestAI | Medium] Train Object Detection AI with 6 lines of code | by Moses Olafenwa | DeepQuestAI | Medium
Or perhaps as was initially suggested, now clearly in the context of utilizing AI/ML/trainingā¦ allow us to dump days of unintentionally recorded vids (FPs) into a folder, v. valid detection vids in another folder containing examples of people on our lawn and in our driveway (v. out of the street walking their dog which would go into FP folder), and irrespective of where the camera is pointed exactly (and thus irrespective of zone settings, and blob detections) if someone (or optionally a neighborās animal) is on/in my yard/driveway irrespective of it being day/night, windy, cloudy, sunny, squirrelsā¦ weād learn of the person/car on my property which i suspect is the 80% case weāre all seeking.
and if youāre not still doing active dev, maybe you could consider addāing one more turn of the crank in the form of a generic extensibility model for swapping in an alternate detector/detection model, which others could develop and plug in? Lots of folks doing cool work in this computer vision space (who are more than likely not the experts you are wrt being the security cam server/manager), who might be motivated to work with your platform and user base to further their work and in the process serving this demand to mutual benefit for all involved.